AdaptiveWeights on a Reading Foundation / V97–V97.1
An AdaptiveWeight unit attached to the frozen reading encoder took the learned reader from 71.4% to 100.0% on the configuration-text formats it read worst, the first AdaptiveWeights result on a Foundation that reads. On usual text styles the attached unit cost 7.5 points, so V97 failed one gate; selecting the unit by decoding confidence in V97.1 did not remove that cost.
SCOPE + LIMITATIONS
- The unit is confidently wrong on many usual-style texts, so its own confidence cannot decide when to attach it; selection needs to know the text’s format family. In the operating path usual styles are read first by exact parsers, so the measured cost falls on texts that would not use the unit. The unit stays out of the default path until a format router exists. One unit, one encoder, one seed.
- This public edition reviews saved evidence; experiments were not rerun for publication.
CONTROL COMPARISON
Attach one rank-8 AdaptiveWeight unit of 294,912 values to the attention and feed-forward input maps of the top 6 of 22 layers of the frozen reading encoder, present only while selected. Train it through the frozen reader head on 3,600 texts in eight weak formats and their usual styles. Evaluate on 800 configurations never used in training: 6,400 weak-format texts and 4,000 usual-style texts, read by the learned reader alone. V97.1 trains nothing: each text is read with the unit detached and attached, and the more confident decoding is kept.
V97 passed detached identity, the weak-format lift (100.0% against 85%) and the frozen-weights check, and failed the usual-style gate at 7.5 points lower against a 1-point limit. V97.1 passed its weak-format and no-training gates and failed the usual-style gate at 86.4% against 91.8%. Both are recorded as failed; the unit is kept as the best of its kind.
- Weak formats · unit detached
- 71.4% of texts fully right
- Weak formats · unit attached
- 100.0%
- Usual styles · unit detached
- 92.8%
- Usual styles · unit attached
- 85.3%
- V97.1 confidence selection · weak formats
- 100.0%; unit chosen for 92.1%
- V97.1 confidence selection · usual styles
- 86.4% against 91.8% required; unit chosen for 58.9%
Question and method
Attach one rank-8 AdaptiveWeight unit of 294,912 values to the attention and feed-forward input maps of the top 6 of 22 layers of the frozen reading encoder, present only while selected. Train it through the frozen reader head on 3,600 texts in eight weak formats and their usual styles. Evaluate on 800 configurations never used in training: 6,400 weak-format texts and 4,000 usual-style texts, read by the learned reader alone. V97.1 trains nothing: each text is read with the unit detached and attached, and the more confident decoding is kept.
Recorded results
| Condition | Recorded result |
|---|---|
| Weak formats · unit detached | 71.4% of texts fully right |
| Weak formats · unit attached | 100.0% |
| Usual styles · unit detached | 92.8% |
| Usual styles · unit attached | 85.3% |
| V97.1 confidence selection · weak formats | 100.0%; unit chosen for 92.1% |
| V97.1 confidence selection · usual styles | 86.4% against 91.8% required; unit chosen for 58.9% |
Comparative standing and locked gates
V97 passed detached identity, the weak-format lift (100.0% against 85%) and the frozen-weights check, and failed the usual-style gate at 7.5 points lower against a 1-point limit. V97.1 passed its weak-format and no-training gates and failed the usual-style gate at 86.4% against 91.8%. Both are recorded as failed; the unit is kept as the best of its kind.
Interpretation and limitations
The unit is confidently wrong on many usual-style texts, so its own confidence cannot decide when to attach it; selection needs to know the text’s format family. In the operating path usual styles are read first by exact parsers, so the measured cost falls on texts that would not use the unit. The unit stays out of the default path until a format router exists. One unit, one encoder, one seed.
SOURCE PROVENANCE
V97: AdaptiveWeights on a Foundation that reads
LABORATORY REPORT / 2026-10-21SOURCE CHECKSUM / SHA-256
72a79afc5ef1fb2dc59a7be1dd41f45f8a971f92f7118b0f2bbd2efaa4c65e73Public journal edition reviewed 2026-10-22. Source documents and saved evidence were inspected; experiments were not rerun for this edition. Proprietary implementation code, model binaries, private infrastructure, and detailed machine records are not published here. Journal identifiers are editorial references. Catalog inclusion does not imply qualification or runtime promotion.